Archetype AI published a new paper demonstrating how an AI model can encode and predict physical behaviours. Newton, the startup's ‘Large Behavior Model’, can learn complex physics principles directly from raw sensor data, without any human guidance.

Significant progress in AI and Robotics this week. Big developments from Archetype AI, NVIDIA, Boston Dynamic, Toyota Research, Google, Adobe, OpenAI, Mistral, Tesla, Meta, and more. Here's everything you need to know and how to make sense out of it:

Nvidia quietly released a new open-sourced, fine-tuned LLM called Llama-3.1-Nemotron-70B-Instruct. On benchmarks, it's outperforming GPT-4o and Claude 3.5 Sonnet, despite being significantly smaller at just 70B parameters

Boston Dynamics and Toyota Research announced a partnership to accelerate the development of humanoids. They plan to create/integrate TRI’s Large Behavior Models (LBMs) into the Atlas electric robot.

Google released a new update to its recently-viral AI Audio Overview feature. Users can now customize AI-generated audio summaries by providing instructions to focus on specific topics or adjusting the expertise level. Huge for auditory learners:

Adobe announced the addition of new video generation capabilities to its Firefly AI model and Premiere Pro The new Firefly Video Model is now in 'limited public beta' and allows users to generate video from text prompts or images

Researchers developed DIAMOND, an AI model that can generate a playable simulation of Counter-Strike at 10 fps within a neural network. The ability to generate interactive, physics-based environments is really interesting here for AI training.

OpenAI introduced Swarm, a new open-source experimental framework designed to simplify the creation and control of multi-agent AI systems. The ability to deploy systems that leverage armies of agents working together is coming fast.

French AI startup Mistral AI just launched two new compact language models The ‘Les Ministraux’ family includes Ministral 3B and Ministral 8B models, and outperforms competitors like Gemma and Llama on benchmarks On-device AI is levelling up, once again

Tesla showed off a new video of Optimus navigating by itself after the Robotaxi event Optimus located charging stations and docked itself, climbed stairs, and autonomously interacted with people

Worldcoin, the ‘proof of personhood’ startup founded by Sam Altman, announced World ID 3.0 to, in part, combat deepfakes. Verifying human identity in the increasing flood of AI-generated content, agents, and systems is clearly going to be important.

Researchers presented a new paper that shows it's possible for robots to learn skills from YouTube without complex video processing. Promising and interesting work:

Meta released Meta Spirit LM, the companies' first open source multimodal model that mixes text and speech The model can essentially understand and create both text and speech in a more natural way, without losing important voice qualities.

We're hiring engineers @Figure_robot: → Systems Integration Engineer → Electrical Engineering (many) → Manufacturing Roles (many) → Controls Eng (many) → Embedded SW → Firmware Eng See all open roles and apply here:

That's it for this week's AI and Robotics breakdown. I share the latest research every week, so follow me @adcock_brett for more. If you found this valuable, consider a like/retweet to spread the word.